Open robotics isn’t gathering steam as fast as AI, but it’s finally making some progress.

NYC — The good news, at UN Open Source Week, is that the experts believe that open robotics is entering a new phase of maturity. This is thanks to open‑source software stacks like Robot Operating System (ROS) 2 and Open Source Computer Vision Library (OpenCV), which now underpin everything from industrial arms to autonomous vehicles. That’s the good news. The bad news is that insiders warn the rollout is still far behind the hype.

As OpenCV creator Gary Bradski stated, despite decades of progress, “right now the robotic smart robotic deployments are disappointing, there aren’t actually many even today,” and that the industry needs shared platforms to stop “everybody right now… reinventing the wheel for themselves.”

Unlike AI, which took off like a rocket with the release of OpenAI’s GPT-3 in 2021, open robotics is moving slowly from lab demos to factory floors and shopping malls. It could be better. Alas, the people building that future say fragmentation, reliability, and funding models still stand between today’s open stacks and truly pervasive “physical AI.”

Kicking off a UN Open Source Week panel on “open robots,” moderator and AWS Managing Director and Head of Financial Services, Beth Fatusin, described a field racing to catch up with cloud and AI in how it shares foundational infrastructure. The robotics ecosystem is “still deeply fragmented with proprietary stacks, closed hardware, incompatible middleware,” she warned, which means “building a robotic system today still means rebuilding things that persons of other teams have already solved behind closed doors.” That duplication, she argued, is “very expensive, and it also limits who gets to participate globally” just as robots begin to enter manufacturing, healthcare, agriculture, and other sectors in earnest.

Against that backdrop, panelists from France, China, and the U.S. sketched a picture of open robotics that is vibrant at the research and component level but still struggling to translate into reliable, scalable deployments.

Bradski, who describes himself as “basically someone who just gets things going,” points to winning the robotic car race, the DARPA Grand Challenge, which became Waymo, to agricultural robots and the first commercial lunar lander, Blue Ghost to show he knows robotics, sadly continued, the economics of critical open components remain broken. “Obviously, I do OpenCV, and it’s probably, you know, saved billions of dollars in software development, and yet you know we get approximately zero back for doing that, and yet have to run this organization.”

That’s because of what he called an “open-source tax” on successful industries: “At some level of scale of an industry, so they’re successful, they’re making millions… They pay some kind of open-source tax, and that’s paid out on the radio song play model. If… a certain open source is used a lot, it gets a proportion of that money, and maybe some is also set aside for new innovation.”

Bradski also highlighted how better open models could blur the line between hardware haves and have‑nots. If “truly capable open-source models” could “get a model and actuate anything,” he said, “hardware can start to be bespoke or produced for certain situations, for instance, even to the level of your particular kitchen, there could be a robot that fits it best.” In that world, you “get some motors or whatever that actuate and then print parts, and you have a robot that’s meant for that environment, or… more appropriate for that society,” with local fabrication also helping “fix and maintain” robots in the field.

From the deployment front, Matthieu Masselin, co-founder and CTO of French startup FOND Robotics, argued that open robotics will only matter if it translates into real‑world systems. “In our view, the most important one in robotics is deployment, and you have to think about impact,” he said. “The only way to have impact in the real world is to deploy actual robots.”

Masselin praised open software like OpenCV and open robot “blueprints that can be built in every university, every college,” but said they “have not been able to crack anything yet in the real world… industrial deployments.” Industrial customers, he pointed out, “don’t want software or hardware; they want a system that someone can stand behind in terms of responsibility,” including reliability, cycle time, and economics. Without clear accountability, “open source… is going to remain limited in its adoption in the real world.”

On AI‑driven robots, Masselin was blunt about the reliability gap between LLMs and physical systems. With language models, “maybe 1% of the time, or even less,” you get “something a little bit crazy,” which is acceptable because “there is a human in the loop that’s able to mediate.” In robotics, by contrast, “if there’s a robot that falls one time every 100 times, this is not something that you can disregard. This can actually be catastrophic.” His conclusion: “Robots will not scale because they are general, they will become general because they scale,” by deploying on well‑defined use cases today, collecting data, and gradually extending into more complex environments.

Panelist Kristine Mo, from Shenzhen‑based humanoid‑robot maker ASPR Robotics, said hardware progress has outpaced intelligence. Her company’s Acrobots—“productivity‑oriented general‑purpose robots” deployed in Chinese malls and parks—already “serve coffee and ice cream on average eight hours a day, making… cups of coffee and ice cream without errors.” But, she said, “we see firsthand how hardware is rapidly advancing, and intelligence is still the bottleneck.”

Mo strongly suggested that open source has to move beyond code dumps to shared, end‑to‑end frameworks. Today, “we’re building this type of technology with different data formats, different model architecture, different simulation environments, and different benchmarks,” making it “incredibly difficult to reproduce someone’s results and compare systems fairly” or “take from a research paper and… deploy into the real world.” Her company’s open platform aims “to connect the dots, to bring the data, model, training, and evaluation and deployment all in one coherent framework,” because “open source should not just open up individual components, but it should give the whole community a shared foundation to build on.”

On global access, she warned that “open source alone is not enough.” Even “the most powerful model in the world” is “in practice still only accessible to a small number of well‑resourced organizations” if it demands “a massive compute cluster and team of specialized engineers.” For researchers and startups, especially in developing economies, “the barrier is not talent, the barrier is access, access to compute, to infrastructures, to toolsets… it is not just about making the code available, but it is genuinely lowering the bar.”

Masselin played down fragmentation among commercial players—“in every industry you have a number of actors, and they all compete, and anyway, this drives innovation… if you look at cars, all cars are different and proprietary”—but said robotics needs “a shared infrastructure” underneath. “When we think about robotics, we think of infrastructure, because we believe it’s going to be relevant to all kinds of industries,” from healthcare to home, he said. The core question, in his view, is “How do you have a shared infrastructure that allows all these fragmented players to compete or be used rather than some conglomerates or monopolies?”

Bradski pointed to data as one possible equalizer. If there is “a way of collecting and sharing data, then it theoretically prevents lock‑out that someone owns everything,” he argued. With enough data, “someone can recreate and retrain a model… to become competitive,” avoiding permanent platform capture. He cited the open‑source driving simulator Carla as a template for sector‑specific, shared accident and incident repositories that could both improve safety and feed new systems.

On the role of governments, the panel converged on a surprisingly pragmatic answer: they should buy robots and fund shared infrastructure. “The number one thing that drives companies is their clients,” Masselin said. “Our governments are probably the largest entities in the world; the number one thing that they can do is to become clients of robotics today, so that they can shape where those companies are going and what kind of products they develop.”

Mo argued that “the most valuable role the government can play is as an ecosystem builder,” investing in “safety standards, data governance framework, shared testing environments, long-term fundamental research support,” which she called “public good.” Fellow panelist Yiming He added that states should both “encourage the regional manufacturing density” and invest in “foundational open source infrastructure or software hardware standards.”

Taken together, the session painted open robotics as standing on a rich foundation—OpenCV‑class software, emerging open platforms, and global manufacturing hubs—but still wrestling with how to fund that commons, harden it for physical reliability, and ensure access beyond a handful of well‑resourced players.

This mirrors broader tensions across open‑source infrastructure: open projects provide the common plumbing, but many of the most visible robots are built by well‑funded companies layering proprietary code and data on top. Advocates like Bradski argue that unless more of the platform itself is commoditized and shared, the field will remain stuck in pilot purgatory with too many custom stacks and not enough reusable, production‑hardened systems.

Their hope is that we’ll take those building blocks and turn them into shared platforms that let robotics companies and researchers stop “reinventing the wheel” and finally move past a world where “there aren’t actually many” truly smart, widely deployed robots. Let’s hope so, for as Bradski observed, we’ll be needing such robots to take care of us in our retirement years.